Week 4 Workshop: The poor image

Can different histories fit the same available image evidence?

17 March 2027

Can different histories fit the same available image evidence?

Learning target. Show that a defined evidence set cannot separate two histories, then specify evidence that could — and accept when the answer stays open.

Before the workshop#

Open the week 4 pack and read observations.csv and tolerance.txt. Ten minutes. Do not open the reveal.

Reading. Hito Steyerl, “In Defense of the Poor Image”, e-flux journal #10 (2009), available free at e-flux.com. Read it for the argument about what circulation does to images. The lecture is explicit about which of its claims are cultural and which would need separate technical evidence.

Last week, and this week#

Week 3 started from a known process and asked how to test its effect. You controlled what you changed.

This week runs the other way. You start from one output and no history, and ask whether the evidence you have can tell you where it came from. It cannot always, and the skill is saying so precisely rather than guessing.

The specimen#

Output 7. Provenance unknown — that is the question. Everything currently known about it is five lines in observations.csv: dimensions, file size, chroma subsampling, and the absence of EXIF.

Two processing histories have been proposed. Route A resizes once and saves once. Route B saves, resizes, then saves again. Both end with the same save, at the same quality, so nothing about the final encode can tell them apart — the only thing they disagree about is whether a full-size JPEG was written before the resize. What each one predicts is supplied, beside what was observed — nothing in this pack would let you derive a file size from a set of steps, and pretending otherwise would make the exercise impossible rather than demanding. Your work is applying the tolerance.

tolerance.txt is the part to read carefully. A route is compatible on file size when its prediction is within 15 KB of the observed size; dimensions, subsampling and EXIF presence must match exactly. If more than one route satisfies that, file size does not distinguish them.

The 15 KB figure is a scenario rule. It is set for this exercise so you can apply a stated threshold consistently. It is not an empirically validated forensic tolerance, and you should not carry it anywhere else.

Indistinguishable is a claim about the evidence, not about the files. Two images that this evidence cannot separate may differ in every byte. Saying “indistinguishable” without saying under what is the error this week exists to prevent.

The scenario is authored throughout. The observations, the predictions and the evidence records were written for this exercise; none of them was produced by running image software, and nothing here supports a claim about how any real tool behaves.

Investigate — 40 minutes#

Test both histories against the observations (15 minutes). For each route, take its predicted values from observations.csv and compare them with the observed values, applying the tolerance. Record the difference per measurement, not just a verdict — “compatible” is a conclusion you show your working for.

Then do the part the pack cannot do for you: sketch a third history that would also fit. Only two were written on cards, and that is a fact about who was asked, not about what is possible.

Choose one further evidence item, and commit first (10 minutes). evidence-menu.csv lists five things you could ask for. Two do not exist, and their absence is marked with a reason. The three that do exist each have a file in evidence/.

Pick one. Then, before you open its file, write down what result would favour route A and what result would favour route B. If you cannot say that in advance, the item cannot discriminate for you — pick another, or record that you could not.

Open only that file, then revise (15 minutes). Each file holds the result of one request and nothing else. Update your conclusion and record what is still unresolved.

Not every available item separates the routes. If yours does not, you have not failed the exercise — say what you asked for, why it was a reasonable request, and what it turned out to be silent about.

Evaluate — 30 minutes#

Write the ambiguity report you are depositing: two route diagrams, your observation table, the scope of indistinguishability, and your evidence request with its justification. Then three paragraphs.

1. State the scope. “These routes are indistinguishable” is incomplete. Finish the sentence: indistinguishable on which measurements, at what tolerance.
2. What does your third route do to your conclusion? You sketched one in the investigation. Say what its existence costs any claim you reached, and whether your chosen evidence would have excluded it too.
3. What is the weakest claim you can defend? If your evidence rules out one route, you have not established another. Write the claim your evidence actually supports, and notice that it is smaller than the one you wanted.

What you should find#

Both routes fit. Dimensions, subsampling and the missing EXIF are identical either way, and the predicted file sizes are 6 KB and 7 KB from the observation, both inside the 15 KB rule.

So the initial evidence does not determine the process, and no amount of staring at it will change that. The question is which further observation would.

One item separates them, and it is worth being precise about why. The retained transformation record names a full-size JPEG as the input to the resize, and the input hash matches a file that was present. Route A resizes the source directly and never has a full-size JPEG to read, so route A is excluded.

Two available items do not separate them, and they fail the same way. The working-directory listing shows a full-size JPEG existed before the output — which looks decisive and is not, because co-existence is not descent. That interim file could be an unrelated export, with the output made straight from the source. The handover note arrives at the same gap from a different direction: it confirms a big version exists somewhere and says nothing about whether it was read.

That distinction — between a record that lists what existed and a record that names what read what — is the whole week in one sentence. Both requests are reasonable, and a report that asks for one, notices the gap and leaves the question open is a correct answer here.

And even having excluded route A, notice what you are left with. Any history that saved at full size before resizing would produce a record of that shape — a save at a different quality, or a route through a third size. Ruling out one route is not establishing another, and the claim you can defend is “not route A, and consistent with route B”.

Deposit#

Both documented branches go into the image collection, with the source and the logs retained. Weeks 10 and 11 come back to this collection: week 10 asks whether a restoration of this output matches the original, and week 11 asks what a provenance record for it would and would not establish.

The five-minute check#

Two route cards fit a supplied thumbnail. Propose one available evidence item that could distinguish them, and explain how.

Assessment#

For The Stemma, this is the reasoning behind its 20% criterion: reconstruct the supported portions and mark the unknown portions as unknown. “Cannot resolve this edge” earns full credit where the evidence does not resolve it, and this week is where you practise writing that sentence so it reads as a finding rather than as a shrug.

For The Chain, bound your conclusion. State the source, the settings and the measure it holds for.